Flash-KMeans: IO-aware exact K-Means, over 200x faster than FAISS on GPUs
Original title:Meet Flash-KMeans: An IO-Aware, Exact K-Means That Runs Over 200× Faster Than FAISS on GPUs
AI summary
A UC Berkeley and UT Austin team has open-sourced Flash-KMeans (Apache 2.0, `pip install flash-kmeans`), an exact implementation of standard Lloyd's k-Means that speeds things up by restructuring the GPU dataflow rather than changing the math or approximating. On NVIDIA H200, it is 17.9x faster end-to-end than the best baseline, 33x faster than cuML and over 200x faster than FAISS. Its FlashAssign kernel avoids materializing the full N×K distance matrix, reducing IO complexity from O(NK) to O(Nd+Kd), with up to 21.2x single-kernel speedup; the Sort-Inverse Update kernel reduces atomic contention by sorting cluster IDs, with up to 6.3x single-kernel speedup. It supports out-of-core processing, completing a single iteration in just 41.4s for 1B data points with K=32768. It is suited to online scenarios such as vector search indexing, sparse attention routing and KV cache compression.
Key Takeaways
- A UC Berkeley and UT Austin team has open-sourced Flash-KMeans (Apache 2.
- 0,`pip install flash-kmeans`),an exact implementation of standard Lloyd's k-Means that speeds…
- On NVIDIA H200,it is 17.
- 9x faster end-to-end than the best baseline,33x faster than cuML and over 200x faster than FAISS.
- Its FlashAssign kernel avoids materializing the full N×K distance matrix,reducing IO complexity…
Quick Facts
- Published
- Updated
- Source
- MarkTechPost(RSS)
- Category
- Products & Apps
- AI score
- 73
Flash-KMeans 把 k-means 从离线预处理拉进了在线循环,200 倍加速不是纸面数字,而是让向量索引重建、稀疏注意力路由这些场景突然可行了。做大规模聚类的可以立刻换掉 FAISS。
Related entities
Source & copyright
This article is based on curated titles and AI summaries from the AIHOT public API, with attribution and links back. Original copyright remains with the original publishers and authors; this site does not mirror full texts.
How to Cite
When citing this article, keep the title, date, and permalink, and credit Orange News as the source.
Orange News: Flash-KMeans: IO-aware exact K-Means, over 200x faster than FAISS on GPUs. June 15, 2026. https://new.nat6.net/en/news/cmqf0xl1601t7slwaf7eww3hm/